Define AVMediaType enum, and use it instead of enum CodecType, which
[libav.git] / libavcodec / twinvq.c
1 /*
2 * TwinVQ decoder
3 * Copyright (c) 2009 Vitor Sessak
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 #include "avcodec.h"
23 #include "get_bits.h"
24 #include "dsputil.h"
25 #include "fft.h"
26 #include "lsp.h"
27
28 #include <math.h>
29 #include <stdint.h>
30
31 #include "twinvq_data.h"
32
33 enum FrameType {
34 FT_SHORT = 0, ///< Short frame (divided in n sub-blocks)
35 FT_MEDIUM, ///< Medium frame (divided in m<n sub-blocks)
36 FT_LONG, ///< Long frame (single sub-block + PPC)
37 FT_PPC, ///< Periodic Peak Component (part of the long frame)
38 };
39
40 /**
41 * Parameters and tables that are different for each frame type
42 */
43 struct FrameMode {
44 uint8_t sub; ///< Number subblocks in each frame
45 const uint16_t *bark_tab;
46
47 /** number of distinct bark scale envelope values */
48 uint8_t bark_env_size;
49
50 const int16_t *bark_cb; ///< codebook for the bark scale envelope (BSE)
51 uint8_t bark_n_coef;///< number of BSE CB coefficients to read
52 uint8_t bark_n_bit; ///< number of bits of the BSE coefs
53
54 //@{
55 /** main codebooks for spectrum data */
56 const int16_t *cb0;
57 const int16_t *cb1;
58 //@}
59
60 uint8_t cb_len_read; ///< number of spectrum coefficients to read
61 };
62
63 /**
64 * Parameters and tables that are different for every combination of
65 * bitrate/sample rate
66 */
67 typedef struct {
68 struct FrameMode fmode[3]; ///< frame type-dependant parameters
69
70 uint16_t size; ///< frame size in samples
71 uint8_t n_lsp; ///< number of lsp coefficients
72 const float *lspcodebook;
73
74 /* number of bits of the different LSP CB coefficients */
75 uint8_t lsp_bit0;
76 uint8_t lsp_bit1;
77 uint8_t lsp_bit2;
78
79 uint8_t lsp_split; ///< number of CB entries for the LSP decoding
80 const int16_t *ppc_shape_cb; ///< PPC shape CB
81
82 /** number of the bits for the PPC period value */
83 uint8_t ppc_period_bit;
84
85 uint8_t ppc_shape_bit; ///< number of bits of the PPC shape CB coeffs
86 uint8_t ppc_shape_len; ///< size of PPC shape CB
87 uint8_t pgain_bit; ///< bits for PPC gain
88
89 /** constant for peak period to peak width conversion */
90 uint16_t peak_per2wid;
91 } ModeTab;
92
93 static const ModeTab mode_08_08 = {
94 {
95 { 8, bark_tab_s08_64, 10, tab.fcb08s , 1, 5, tab.cb0808s0, tab.cb0808s1, 18},
96 { 2, bark_tab_m08_256, 20, tab.fcb08m , 2, 5, tab.cb0808m0, tab.cb0808m1, 16},
97 { 1, bark_tab_l08_512, 30, tab.fcb08l , 3, 6, tab.cb0808l0, tab.cb0808l1, 17}
98 },
99 512 , 12, tab.lsp08, 1, 5, 3, 3, tab.shape08 , 8, 28, 20, 6, 40
100 };
101
102 static const ModeTab mode_11_08 = {
103 {
104 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1108s0, tab.cb1108s1, 29},
105 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1108m0, tab.cb1108m1, 24},
106 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1108l0, tab.cb1108l1, 27}
107 },
108 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
109 };
110
111 static const ModeTab mode_11_10 = {
112 {
113 { 8, bark_tab_s11_64, 10, tab.fcb11s , 1, 5, tab.cb1110s0, tab.cb1110s1, 21},
114 { 2, bark_tab_m11_256, 20, tab.fcb11m , 2, 5, tab.cb1110m0, tab.cb1110m1, 18},
115 { 1, bark_tab_l11_512, 30, tab.fcb11l , 3, 6, tab.cb1110l0, tab.cb1110l1, 20}
116 },
117 512 , 16, tab.lsp11, 1, 6, 4, 3, tab.shape11 , 9, 36, 30, 7, 90
118 };
119
120 static const ModeTab mode_16_16 = {
121 {
122 { 8, bark_tab_s16_128, 10, tab.fcb16s , 1, 5, tab.cb1616s0, tab.cb1616s1, 16},
123 { 2, bark_tab_m16_512, 20, tab.fcb16m , 2, 5, tab.cb1616m0, tab.cb1616m1, 15},
124 { 1, bark_tab_l16_1024,30, tab.fcb16l , 3, 6, tab.cb1616l0, tab.cb1616l1, 16}
125 },
126 1024, 16, tab.lsp16, 1, 6, 4, 3, tab.shape16 , 9, 56, 60, 7, 180
127 };
128
129 static const ModeTab mode_22_20 = {
130 {
131 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2220s0, tab.cb2220s1, 18},
132 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2220m0, tab.cb2220m1, 17},
133 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2220l0, tab.cb2220l1, 18}
134 },
135 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
136 };
137
138 static const ModeTab mode_22_24 = {
139 {
140 { 8, bark_tab_s22_128, 10, tab.fcb22s_1, 1, 6, tab.cb2224s0, tab.cb2224s1, 15},
141 { 2, bark_tab_m22_512, 20, tab.fcb22m_1, 2, 6, tab.cb2224m0, tab.cb2224m1, 14},
142 { 1, bark_tab_l22_1024,32, tab.fcb22l_1, 4, 6, tab.cb2224l0, tab.cb2224l1, 15}
143 },
144 1024, 16, tab.lsp22_1, 1, 6, 4, 3, tab.shape22_1, 9, 56, 36, 7, 144
145 };
146
147 static const ModeTab mode_22_32 = {
148 {
149 { 4, bark_tab_s22_128, 10, tab.fcb22s_2, 1, 6, tab.cb2232s0, tab.cb2232s1, 11},
150 { 2, bark_tab_m22_256, 20, tab.fcb22m_2, 2, 6, tab.cb2232m0, tab.cb2232m1, 11},
151 { 1, bark_tab_l22_512, 32, tab.fcb22l_2, 4, 6, tab.cb2232l0, tab.cb2232l1, 12}
152 },
153 512 , 16, tab.lsp22_2, 1, 6, 4, 4, tab.shape22_2, 9, 56, 36, 7, 72
154 };
155
156 static const ModeTab mode_44_40 = {
157 {
158 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4440s0, tab.cb4440s1, 18},
159 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4440m0, tab.cb4440m1, 17},
160 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4440l0, tab.cb4440l1, 17}
161 },
162 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
163 };
164
165 static const ModeTab mode_44_48 = {
166 {
167 {16, bark_tab_s44_128, 10, tab.fcb44s , 1, 6, tab.cb4448s0, tab.cb4448s1, 15},
168 { 4, bark_tab_m44_512, 20, tab.fcb44m , 2, 6, tab.cb4448m0, tab.cb4448m1, 14},
169 { 1, bark_tab_l44_2048,40, tab.fcb44l , 4, 6, tab.cb4448l0, tab.cb4448l1, 14}
170 },
171 2048, 20, tab.lsp44, 1, 6, 4, 4, tab.shape44 , 9, 84, 54, 7, 432
172 };
173
174 typedef struct TwinContext {
175 AVCodecContext *avctx;
176 DSPContext dsp;
177 FFTContext mdct_ctx[3];
178
179 const ModeTab *mtab;
180
181 // history
182 float lsp_hist[2][20]; ///< LSP coefficients of the last frame
183 float bark_hist[3][2][40]; ///< BSE coefficients of last frame
184
185 // bitstream parameters
186 int16_t permut[4][4096];
187 uint8_t length[4][2]; ///< main codebook stride
188 uint8_t length_change[4];
189 uint8_t bits_main_spec[2][4][2]; ///< bits for the main codebook
190 int bits_main_spec_change[4];
191 int n_div[4];
192
193 float *spectrum;
194 float *curr_frame; ///< non-interleaved output
195 float *prev_frame; ///< non-interleaved previous frame
196 int last_block_pos[2];
197
198 float *cos_tabs[3];
199
200 // scratch buffers
201 float *tmp_buf;
202 } TwinContext;
203
204 #define PPC_SHAPE_CB_SIZE 64
205 #define SUB_AMP_MAX 4500.0
206 #define MULAW_MU 100.0
207 #define GAIN_BITS 8
208 #define AMP_MAX 13000.0
209 #define SUB_GAIN_BITS 5
210 #define WINDOW_TYPE_BITS 4
211 #define PGAIN_MU 200
212
213 /** @note not speed critical, hence not optimized */
214 static void memset_float(float *buf, float val, int size)
215 {
216 while (size--)
217 *buf++ = val;
218 }
219
220 /**
221 * Evaluate a single LPC amplitude spectrum envelope coefficient from the line
222 * spectrum pairs.
223 *
224 * @param lsp a vector of the cosinus of the LSP values
225 * @param cos_val cos(PI*i/N) where i is the index of the LPC amplitude
226 * @param order the order of the LSP (and the size of the *lsp buffer). Must
227 * be a multiple of four.
228 * @return the LPC value
229 *
230 * @todo reuse code from vorbis_dec.c: vorbis_floor0_decode
231 */
232 static float eval_lpc_spectrum(const float *lsp, float cos_val, int order)
233 {
234 int j;
235 float p = 0.5f;
236 float q = 0.5f;
237 float two_cos_w = 2.0f*cos_val;
238
239 for (j = 0; j + 1 < order; j += 2*2) {
240 // Unroll the loop once since order is a multiple of four
241 q *= lsp[j ] - two_cos_w;
242 p *= lsp[j+1] - two_cos_w;
243
244 q *= lsp[j+2] - two_cos_w;
245 p *= lsp[j+3] - two_cos_w;
246 }
247
248 p *= p * (2.0f - two_cos_w);
249 q *= q * (2.0f + two_cos_w);
250
251 return 0.5 / (p + q);
252 }
253
254 /**
255 * Evaluates the LPC amplitude spectrum envelope from the line spectrum pairs.
256 */
257 static void eval_lpcenv(TwinContext *tctx, const float *cos_vals, float *lpc)
258 {
259 int i;
260 const ModeTab *mtab = tctx->mtab;
261 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
262
263 for (i = 0; i < size_s/2; i++) {
264 float cos_i = tctx->cos_tabs[0][i];
265 lpc[i] = eval_lpc_spectrum(cos_vals, cos_i, mtab->n_lsp);
266 lpc[size_s-i-1] = eval_lpc_spectrum(cos_vals, -cos_i, mtab->n_lsp);
267 }
268 }
269
270 static void interpolate(float *out, float v1, float v2, int size)
271 {
272 int i;
273 float step = (v1 - v2)/(size + 1);
274
275 for (i = 0; i < size; i++) {
276 v2 += step;
277 out[i] = v2;
278 }
279 }
280
281 static inline float get_cos(int idx, int part, const float *cos_tab, int size)
282 {
283 return part ? -cos_tab[size - idx - 1] :
284 cos_tab[ idx ];
285 }
286
287 /**
288 * Evaluates the LPC amplitude spectrum envelope from the line spectrum pairs.
289 * Probably for speed reasons, the coefficients are evaluated as
290 * siiiibiiiisiiiibiiiisiiiibiiiisiiiibiiiis ...
291 * where s is an evaluated value, i is a value interpolated from the others
292 * and b might be either calculated or interpolated, depending on an
293 * unexplained condition.
294 *
295 * @param step the size of a block "siiiibiiii"
296 * @param in the cosinus of the LSP data
297 * @param part is 0 for 0...PI (positive cossinus values) and 1 for PI...2PI
298 (negative cossinus values)
299 * @param size the size of the whole output
300 */
301 static inline void eval_lpcenv_or_interp(TwinContext *tctx,
302 enum FrameType ftype,
303 float *out, const float *in,
304 int size, int step, int part)
305 {
306 int i;
307 const ModeTab *mtab = tctx->mtab;
308 const float *cos_tab = tctx->cos_tabs[ftype];
309
310 // Fill the 's'
311 for (i = 0; i < size; i += step)
312 out[i] =
313 eval_lpc_spectrum(in,
314 get_cos(i, part, cos_tab, size),
315 mtab->n_lsp);
316
317 // Fill the 'iiiibiiii'
318 for (i = step; i <= size - 2*step; i += step) {
319 if (out[i + step] + out[i - step] > 1.95*out[i] ||
320 out[i + step] >= out[i - step]) {
321 interpolate(out + i - step + 1, out[i], out[i-step], step - 1);
322 } else {
323 out[i - step/2] =
324 eval_lpc_spectrum(in,
325 get_cos(i-step/2, part, cos_tab, size),
326 mtab->n_lsp);
327 interpolate(out + i - step + 1, out[i-step/2], out[i-step ], step/2 - 1);
328 interpolate(out + i - step/2 + 1, out[i ], out[i-step/2], step/2 - 1);
329 }
330 }
331
332 interpolate(out + size - 2*step + 1, out[size-step], out[size - 2*step], step - 1);
333 }
334
335 static void eval_lpcenv_2parts(TwinContext *tctx, enum FrameType ftype,
336 const float *buf, float *lpc,
337 int size, int step)
338 {
339 eval_lpcenv_or_interp(tctx, ftype, lpc , buf, size/2, step, 0);
340 eval_lpcenv_or_interp(tctx, ftype, lpc + size/2, buf, size/2, 2*step, 1);
341
342 interpolate(lpc+size/2-step+1, lpc[size/2], lpc[size/2-step], step);
343
344 memset_float(lpc + size - 2*step + 1, lpc[size - 2*step], 2*step - 1);
345 }
346
347 /**
348 * Inverse quantization. Read CB coefficients for cb1 and cb2 from the
349 * bitstream, sum the corresponding vectors and write the result to *out
350 * after permutation.
351 */
352 static void dequant(TwinContext *tctx, GetBitContext *gb, float *out,
353 enum FrameType ftype,
354 const int16_t *cb0, const int16_t *cb1, int cb_len)
355 {
356 int pos = 0;
357 int i, j;
358
359 for (i = 0; i < tctx->n_div[ftype]; i++) {
360 int tmp0, tmp1;
361 int sign0 = 1;
362 int sign1 = 1;
363 const int16_t *tab0, *tab1;
364 int length = tctx->length[ftype][i >= tctx->length_change[ftype]];
365 int bitstream_second_part = (i >= tctx->bits_main_spec_change[ftype]);
366
367 int bits = tctx->bits_main_spec[0][ftype][bitstream_second_part];
368 if (bits == 7) {
369 if (get_bits1(gb))
370 sign0 = -1;
371 bits = 6;
372 }
373 tmp0 = get_bits(gb, bits);
374
375 bits = tctx->bits_main_spec[1][ftype][bitstream_second_part];
376
377 if (bits == 7) {
378 if (get_bits1(gb))
379 sign1 = -1;
380
381 bits = 6;
382 }
383 tmp1 = get_bits(gb, bits);
384
385 tab0 = cb0 + tmp0*cb_len;
386 tab1 = cb1 + tmp1*cb_len;
387
388 for (j = 0; j < length; j++)
389 out[tctx->permut[ftype][pos+j]] = sign0*tab0[j] + sign1*tab1[j];
390
391 pos += length;
392 }
393
394 }
395
396 static inline float mulawinv(float y, float clip, float mu)
397 {
398 y = av_clipf(y/clip, -1, 1);
399 return clip * FFSIGN(y) * (exp(log(1+mu) * fabs(y)) - 1) / mu;
400 }
401
402 /**
403 * Evaluate a*b/400 rounded to the nearest integer. When, for example,
404 * a*b == 200 and the nearest integer is ill-defined, use a table to emulate
405 * the following broken float-based implementation used by the binary decoder:
406 *
407 * \code
408 * static int very_broken_op(int a, int b)
409 * {
410 * static float test; // Ugh, force gcc to do the division first...
411 *
412 * test = a/400.;
413 * return b * test + 0.5;
414 * }
415 * \endcode
416 *
417 * @note if this function is replaced by just ROUNDED_DIV(a*b,400.), the stddev
418 * between the original file (before encoding with Yamaha encoder) and the
419 * decoded output increases, which leads one to believe that the encoder expects
420 * exactly this broken calculation.
421 */
422 static int very_broken_op(int a, int b)
423 {
424 int x = a*b + 200;
425 int size;
426 const uint8_t *rtab;
427
428 if (x%400 || b%5)
429 return x/400;
430
431 x /= 400;
432
433 size = tabs[b/5].size;
434 rtab = tabs[b/5].tab;
435 return x - rtab[size*av_log2(2*(x - 1)/size)+(x - 1)%size];
436 }
437
438 /**
439 * Sum to data a periodic peak of a given period, width and shape.
440 *
441 * @param period the period of the peak divised by 400.0
442 */
443 static void add_peak(int period, int width, const float *shape,
444 float ppc_gain, float *speech, int len)
445 {
446 int i, j;
447
448 const float *shape_end = shape + len;
449 int center;
450
451 // First peak centered around zero
452 for (i = 0; i < width/2; i++)
453 speech[i] += ppc_gain * *shape++;
454
455 for (i = 1; i < ROUNDED_DIV(len,width) ; i++) {
456 center = very_broken_op(period, i);
457 for (j = -width/2; j < (width+1)/2; j++)
458 speech[j+center] += ppc_gain * *shape++;
459 }
460
461 // For the last block, be careful not to go beyond the end of the buffer
462 center = very_broken_op(period, i);
463 for (j = -width/2; j < (width + 1)/2 && shape < shape_end; j++)
464 speech[j+center] += ppc_gain * *shape++;
465 }
466
467 static void decode_ppc(TwinContext *tctx, int period_coef, const float *shape,
468 float ppc_gain, float *speech)
469 {
470 const ModeTab *mtab = tctx->mtab;
471 int isampf = tctx->avctx->sample_rate/1000;
472 int ibps = tctx->avctx->bit_rate/(1000 * tctx->avctx->channels);
473 int min_period = ROUNDED_DIV( 40*2*mtab->size, isampf);
474 int max_period = ROUNDED_DIV(6*40*2*mtab->size, isampf);
475 int period_range = max_period - min_period;
476
477 // This is actually the period multiplied by 400. It is just linearly coded
478 // between its maximum and minimum value.
479 int period = min_period +
480 ROUNDED_DIV(period_coef*period_range, (1 << mtab->ppc_period_bit) - 1);
481 int width;
482
483 if (isampf == 22 && ibps == 32) {
484 // For some unknown reason, NTT decided to code this case differently...
485 width = ROUNDED_DIV((period + 800)* mtab->peak_per2wid, 400*mtab->size);
486 } else
487 width = (period )* mtab->peak_per2wid/(400*mtab->size);
488
489 add_peak(period, width, shape, ppc_gain, speech, mtab->ppc_shape_len);
490 }
491
492 static void dec_gain(TwinContext *tctx, GetBitContext *gb, enum FrameType ftype,
493 float *out)
494 {
495 const ModeTab *mtab = tctx->mtab;
496 int i, j;
497 int sub = mtab->fmode[ftype].sub;
498 float step = AMP_MAX / ((1 << GAIN_BITS) - 1);
499 float sub_step = SUB_AMP_MAX / ((1 << SUB_GAIN_BITS) - 1);
500
501 if (ftype == FT_LONG) {
502 for (i = 0; i < tctx->avctx->channels; i++)
503 out[i] = (1./(1<<13)) *
504 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
505 AMP_MAX, MULAW_MU);
506 } else {
507 for (i = 0; i < tctx->avctx->channels; i++) {
508 float val = (1./(1<<23)) *
509 mulawinv(step * 0.5 + step * get_bits(gb, GAIN_BITS),
510 AMP_MAX, MULAW_MU);
511
512 for (j = 0; j < sub; j++) {
513 out[i*sub + j] =
514 val*mulawinv(sub_step* 0.5 +
515 sub_step* get_bits(gb, SUB_GAIN_BITS),
516 SUB_AMP_MAX, MULAW_MU);
517 }
518 }
519 }
520 }
521
522 /**
523 * Rearrange the LSP coefficients so that they have a minimum distance of
524 * min_dist. This function does it exactly as described in section of 3.2.4
525 * of the G.729 specification (but interestingly is different from what the
526 * reference decoder actually does).
527 */
528 static void rearrange_lsp(int order, float *lsp, float min_dist)
529 {
530 int i;
531 float min_dist2 = min_dist * 0.5;
532 for (i = 1; i < order; i++)
533 if (lsp[i] - lsp[i-1] < min_dist) {
534 float avg = (lsp[i] + lsp[i-1]) * 0.5;
535
536 lsp[i-1] = avg - min_dist2;
537 lsp[i ] = avg + min_dist2;
538 }
539 }
540
541 static void decode_lsp(TwinContext *tctx, int lpc_idx1, uint8_t *lpc_idx2,
542 int lpc_hist_idx, float *lsp, float *hist)
543 {
544 const ModeTab *mtab = tctx->mtab;
545 int i, j;
546
547 const float *cb = mtab->lspcodebook;
548 const float *cb2 = cb + (1 << mtab->lsp_bit1)*mtab->n_lsp;
549 const float *cb3 = cb2 + (1 << mtab->lsp_bit2)*mtab->n_lsp;
550
551 const int8_t funny_rounding[4] = {
552 -2,
553 mtab->lsp_split == 4 ? -2 : 1,
554 mtab->lsp_split == 4 ? -2 : 1,
555 0
556 };
557
558 j = 0;
559 for (i = 0; i < mtab->lsp_split; i++) {
560 int chunk_end = ((i + 1)*mtab->n_lsp + funny_rounding[i])/mtab->lsp_split;
561 for (; j < chunk_end; j++)
562 lsp[j] = cb [lpc_idx1 * mtab->n_lsp + j] +
563 cb2[lpc_idx2[i] * mtab->n_lsp + j];
564 }
565
566 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
567
568 for (i = 0; i < mtab->n_lsp; i++) {
569 float tmp1 = 1. - cb3[lpc_hist_idx*mtab->n_lsp + i];
570 float tmp2 = hist[i] * cb3[lpc_hist_idx*mtab->n_lsp + i];
571 hist[i] = lsp[i];
572 lsp[i] = lsp[i] * tmp1 + tmp2;
573 }
574
575 rearrange_lsp(mtab->n_lsp, lsp, 0.0001);
576 rearrange_lsp(mtab->n_lsp, lsp, 0.000095);
577 ff_sort_nearly_sorted_floats(lsp, mtab->n_lsp);
578 }
579
580 static void dec_lpc_spectrum_inv(TwinContext *tctx, float *lsp,
581 enum FrameType ftype, float *lpc)
582 {
583 int i;
584 int size = tctx->mtab->size / tctx->mtab->fmode[ftype].sub;
585
586 for (i = 0; i < tctx->mtab->n_lsp; i++)
587 lsp[i] = 2*cos(lsp[i]);
588
589 switch (ftype) {
590 case FT_LONG:
591 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 8);
592 break;
593 case FT_MEDIUM:
594 eval_lpcenv_2parts(tctx, ftype, lsp, lpc, size, 2);
595 break;
596 case FT_SHORT:
597 eval_lpcenv(tctx, lsp, lpc);
598 break;
599 }
600 }
601
602 static void imdct_and_window(TwinContext *tctx, enum FrameType ftype, int wtype,
603 float *in, float *prev, int ch)
604 {
605 const ModeTab *mtab = tctx->mtab;
606 int bsize = mtab->size / mtab->fmode[ftype].sub;
607 int size = mtab->size;
608 float *buf1 = tctx->tmp_buf;
609 int j;
610 int wsize; // Window size
611 float *out = tctx->curr_frame + 2*ch*mtab->size;
612 float *out2 = out;
613 float *prev_buf;
614 int first_wsize;
615
616 static const uint8_t wtype_to_wsize[] = {0, 0, 2, 2, 2, 1, 0, 1, 1};
617 int types_sizes[] = {
618 mtab->size / mtab->fmode[FT_LONG ].sub,
619 mtab->size / mtab->fmode[FT_MEDIUM].sub,
620 mtab->size / (2*mtab->fmode[FT_SHORT ].sub),
621 };
622
623 wsize = types_sizes[wtype_to_wsize[wtype]];
624 first_wsize = wsize;
625 prev_buf = prev + (size - bsize)/2;
626
627 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
628 int sub_wtype = ftype == FT_MEDIUM ? 8 : wtype;
629
630 if (!j && wtype == 4)
631 sub_wtype = 4;
632 else if (j == mtab->fmode[ftype].sub-1 && wtype == 7)
633 sub_wtype = 7;
634
635 wsize = types_sizes[wtype_to_wsize[sub_wtype]];
636
637 ff_imdct_half(&tctx->mdct_ctx[ftype], buf1 + bsize*j, in + bsize*j);
638
639 tctx->dsp.vector_fmul_window(out2,
640 prev_buf + (bsize-wsize)/2,
641 buf1 + bsize*j,
642 ff_sine_windows[av_log2(wsize)],
643 0.0,
644 wsize/2);
645 out2 += wsize;
646
647 memcpy(out2, buf1 + bsize*j + wsize/2, (bsize - wsize/2)*sizeof(float));
648
649 out2 += ftype == FT_MEDIUM ? (bsize-wsize)/2 : bsize - wsize;
650
651 prev_buf = buf1 + bsize*j + bsize/2;
652 }
653
654 tctx->last_block_pos[ch] = (size + first_wsize)/2;
655 }
656
657 static void imdct_output(TwinContext *tctx, enum FrameType ftype, int wtype,
658 float *out)
659 {
660 const ModeTab *mtab = tctx->mtab;
661 float *prev_buf = tctx->prev_frame + tctx->last_block_pos[0];
662 int i, j;
663
664 for (i = 0; i < tctx->avctx->channels; i++) {
665 imdct_and_window(tctx, ftype, wtype,
666 tctx->spectrum + i*mtab->size,
667 prev_buf + 2*i*mtab->size,
668 i);
669 }
670
671 if (tctx->avctx->channels == 2) {
672 for (i = 0; i < mtab->size - tctx->last_block_pos[0]; i++) {
673 float f1 = prev_buf[ i];
674 float f2 = prev_buf[2*mtab->size + i];
675 out[2*i ] = f1 + f2;
676 out[2*i + 1] = f1 - f2;
677 }
678 for (j = 0; i < mtab->size; j++,i++) {
679 float f1 = tctx->curr_frame[ j];
680 float f2 = tctx->curr_frame[2*mtab->size + j];
681 out[2*i ] = f1 + f2;
682 out[2*i + 1] = f1 - f2;
683 }
684 } else {
685 memcpy(out, prev_buf,
686 (mtab->size - tctx->last_block_pos[0]) * sizeof(*out));
687
688 out += mtab->size - tctx->last_block_pos[0];
689
690 memcpy(out, tctx->curr_frame,
691 (tctx->last_block_pos[0]) * sizeof(*out));
692 }
693
694 }
695
696 static void dec_bark_env(TwinContext *tctx, const uint8_t *in, int use_hist,
697 int ch, float *out, float gain, enum FrameType ftype)
698 {
699 const ModeTab *mtab = tctx->mtab;
700 int i,j;
701 float *hist = tctx->bark_hist[ftype][ch];
702 float val = ((const float []) {0.4, 0.35, 0.28})[ftype];
703 int bark_n_coef = mtab->fmode[ftype].bark_n_coef;
704 int fw_cb_len = mtab->fmode[ftype].bark_env_size / bark_n_coef;
705 int idx = 0;
706
707 for (i = 0; i < fw_cb_len; i++)
708 for (j = 0; j < bark_n_coef; j++, idx++) {
709 float tmp2 =
710 mtab->fmode[ftype].bark_cb[fw_cb_len*in[j] + i] * (1./4096);
711 float st = use_hist ?
712 (1. - val) * tmp2 + val*hist[idx] + 1. : tmp2 + 1.;
713
714 hist[idx] = tmp2;
715 if (st < -1.) st = 1.;
716
717 memset_float(out, st * gain, mtab->fmode[ftype].bark_tab[idx]);
718 out += mtab->fmode[ftype].bark_tab[idx];
719 }
720
721 }
722
723 static void read_and_decode_spectrum(TwinContext *tctx, GetBitContext *gb,
724 float *out, enum FrameType ftype)
725 {
726 const ModeTab *mtab = tctx->mtab;
727 int channels = tctx->avctx->channels;
728 int sub = mtab->fmode[ftype].sub;
729 int block_size = mtab->size / sub;
730 float gain[channels*sub];
731 float ppc_shape[mtab->ppc_shape_len * channels * 4];
732 uint8_t bark1[channels][sub][mtab->fmode[ftype].bark_n_coef];
733 uint8_t bark_use_hist[channels][sub];
734
735 uint8_t lpc_idx1[channels];
736 uint8_t lpc_idx2[channels][tctx->mtab->lsp_split];
737 uint8_t lpc_hist_idx[channels];
738
739 int i, j, k;
740
741 dequant(tctx, gb, out, ftype,
742 mtab->fmode[ftype].cb0, mtab->fmode[ftype].cb1,
743 mtab->fmode[ftype].cb_len_read);
744
745 for (i = 0; i < channels; i++)
746 for (j = 0; j < sub; j++)
747 for (k = 0; k < mtab->fmode[ftype].bark_n_coef; k++)
748 bark1[i][j][k] =
749 get_bits(gb, mtab->fmode[ftype].bark_n_bit);
750
751 for (i = 0; i < channels; i++)
752 for (j = 0; j < sub; j++)
753 bark_use_hist[i][j] = get_bits1(gb);
754
755 dec_gain(tctx, gb, ftype, gain);
756
757 for (i = 0; i < channels; i++) {
758 lpc_hist_idx[i] = get_bits(gb, tctx->mtab->lsp_bit0);
759 lpc_idx1 [i] = get_bits(gb, tctx->mtab->lsp_bit1);
760
761 for (j = 0; j < tctx->mtab->lsp_split; j++)
762 lpc_idx2[i][j] = get_bits(gb, tctx->mtab->lsp_bit2);
763 }
764
765 if (ftype == FT_LONG) {
766 int cb_len_p = (tctx->n_div[3] + mtab->ppc_shape_len*channels - 1)/
767 tctx->n_div[3];
768 dequant(tctx, gb, ppc_shape, FT_PPC, mtab->ppc_shape_cb,
769 mtab->ppc_shape_cb + cb_len_p*PPC_SHAPE_CB_SIZE, cb_len_p);
770 }
771
772 for (i = 0; i < channels; i++) {
773 float *chunk = out + mtab->size * i;
774 float lsp[tctx->mtab->n_lsp];
775
776 for (j = 0; j < sub; j++) {
777 dec_bark_env(tctx, bark1[i][j], bark_use_hist[i][j], i,
778 tctx->tmp_buf, gain[sub*i+j], ftype);
779
780 tctx->dsp.vector_fmul(chunk + block_size*j, tctx->tmp_buf,
781 block_size);
782
783 }
784
785 if (ftype == FT_LONG) {
786 float pgain_step = 25000. / ((1 << mtab->pgain_bit) - 1);
787 int p_coef = get_bits(gb, tctx->mtab->ppc_period_bit);
788 int g_coef = get_bits(gb, tctx->mtab->pgain_bit);
789 float v = 1./8192*
790 mulawinv(pgain_step*g_coef+ pgain_step/2, 25000., PGAIN_MU);
791
792 decode_ppc(tctx, p_coef, ppc_shape + i*mtab->ppc_shape_len, v,
793 chunk);
794 }
795
796 decode_lsp(tctx, lpc_idx1[i], lpc_idx2[i], lpc_hist_idx[i], lsp,
797 tctx->lsp_hist[i]);
798
799 dec_lpc_spectrum_inv(tctx, lsp, ftype, tctx->tmp_buf);
800
801 for (j = 0; j < mtab->fmode[ftype].sub; j++) {
802 tctx->dsp.vector_fmul(chunk, tctx->tmp_buf, block_size);
803 chunk += block_size;
804 }
805 }
806 }
807
808 static int twin_decode_frame(AVCodecContext * avctx, void *data,
809 int *data_size, AVPacket *avpkt)
810 {
811 const uint8_t *buf = avpkt->data;
812 int buf_size = avpkt->size;
813 TwinContext *tctx = avctx->priv_data;
814 GetBitContext gb;
815 const ModeTab *mtab = tctx->mtab;
816 float *out = data;
817 enum FrameType ftype;
818 int window_type;
819 static const enum FrameType wtype_to_ftype_table[] = {
820 FT_LONG, FT_LONG, FT_SHORT, FT_LONG,
821 FT_MEDIUM, FT_LONG, FT_LONG, FT_MEDIUM, FT_MEDIUM
822 };
823
824 if (buf_size*8 < avctx->bit_rate*mtab->size/avctx->sample_rate + 8) {
825 av_log(avctx, AV_LOG_ERROR,
826 "Frame too small (%d bytes). Truncated file?\n", buf_size);
827 *data_size = 0;
828 return buf_size;
829 }
830
831 init_get_bits(&gb, buf, buf_size * 8);
832 skip_bits(&gb, get_bits(&gb, 8));
833 window_type = get_bits(&gb, WINDOW_TYPE_BITS);
834
835 if (window_type > 8) {
836 av_log(avctx, AV_LOG_ERROR, "Invalid window type, broken sample?\n");
837 return -1;
838 }
839
840 ftype = wtype_to_ftype_table[window_type];
841
842 read_and_decode_spectrum(tctx, &gb, tctx->spectrum, ftype);
843
844 imdct_output(tctx, ftype, window_type, out);
845
846 FFSWAP(float*, tctx->curr_frame, tctx->prev_frame);
847
848 if (tctx->avctx->frame_number < 2) {
849 *data_size=0;
850 return buf_size;
851 }
852
853 tctx->dsp.vector_clipf(out, out, -32700./(1<<15), 32700./(1<<15),
854 avctx->channels * mtab->size);
855
856 *data_size = mtab->size*avctx->channels*4;
857
858 return buf_size;
859 }
860
861 /**
862 * Init IMDCT and windowing tables
863 */
864 static av_cold void init_mdct_win(TwinContext *tctx)
865 {
866 int i,j;
867 const ModeTab *mtab = tctx->mtab;
868 int size_s = mtab->size / mtab->fmode[FT_SHORT].sub;
869 int size_m = mtab->size / mtab->fmode[FT_MEDIUM].sub;
870 int channels = tctx->avctx->channels;
871 float norm = channels == 1 ? 2. : 1.;
872
873 for (i = 0; i < 3; i++) {
874 int bsize = tctx->mtab->size/tctx->mtab->fmode[i].sub;
875 ff_mdct_init(&tctx->mdct_ctx[i], av_log2(bsize) + 1, 1,
876 -sqrt(norm/bsize) / (1<<15));
877 }
878
879 tctx->tmp_buf = av_malloc(mtab->size * sizeof(*tctx->tmp_buf));
880
881 tctx->spectrum = av_malloc(2*mtab->size*channels*sizeof(float));
882 tctx->curr_frame = av_malloc(2*mtab->size*channels*sizeof(float));
883 tctx->prev_frame = av_malloc(2*mtab->size*channels*sizeof(float));
884
885 for (i = 0; i < 3; i++) {
886 int m = 4*mtab->size/mtab->fmode[i].sub;
887 double freq = 2*M_PI/m;
888 tctx->cos_tabs[i] = av_malloc((m/4)*sizeof(*tctx->cos_tabs));
889
890 for (j = 0; j <= m/8; j++)
891 tctx->cos_tabs[i][j] = cos((2*j + 1)*freq);
892 for (j = 1; j < m/8; j++)
893 tctx->cos_tabs[i][m/4-j] = tctx->cos_tabs[i][j];
894 }
895
896
897 ff_init_ff_sine_windows(av_log2(size_m));
898 ff_init_ff_sine_windows(av_log2(size_s/2));
899 ff_init_ff_sine_windows(av_log2(mtab->size));
900 }
901
902 /**
903 * Interpret the data as if it were a num_blocks x line_len[0] matrix and for
904 * each line do a cyclic permutation, i.e.
905 * abcdefghijklm -> defghijklmabc
906 * where the amount to be shifted is evaluated depending on the column.
907 */
908 static void permutate_in_line(int16_t *tab, int num_vect, int num_blocks,
909 int block_size,
910 const uint8_t line_len[2], int length_div,
911 enum FrameType ftype)
912
913 {
914 int i,j;
915
916 for (i = 0; i < line_len[0]; i++) {
917 int shift;
918
919 if (num_blocks == 1 ||
920 (ftype == FT_LONG && num_vect % num_blocks) ||
921 (ftype != FT_LONG && num_vect & 1 ) ||
922 i == line_len[1]) {
923 shift = 0;
924 } else if (ftype == FT_LONG) {
925 shift = i;
926 } else
927 shift = i*i;
928
929 for (j = 0; j < num_vect && (j+num_vect*i < block_size*num_blocks); j++)
930 tab[i*num_vect+j] = i*num_vect + (j + shift) % num_vect;
931 }
932 }
933
934 /**
935 * Interpret the input data as in the following table:
936 *
937 * \verbatim
938 *
939 * abcdefgh
940 * ijklmnop
941 * qrstuvw
942 * x123456
943 *
944 * \endverbatim
945 *
946 * and transpose it, giving the output
947 * aiqxbjr1cks2dlt3emu4fvn5gow6hp
948 */
949 static void transpose_perm(int16_t *out, int16_t *in, int num_vect,
950 const uint8_t line_len[2], int length_div)
951 {
952 int i,j;
953 int cont= 0;
954 for (i = 0; i < num_vect; i++)
955 for (j = 0; j < line_len[i >= length_div]; j++)
956 out[cont++] = in[j*num_vect + i];
957 }
958
959 static void linear_perm(int16_t *out, int16_t *in, int n_blocks, int size)
960 {
961 int block_size = size/n_blocks;
962 int i;
963
964 for (i = 0; i < size; i++)
965 out[i] = block_size * (in[i] % n_blocks) + in[i] / n_blocks;
966 }
967
968 static av_cold void construct_perm_table(TwinContext *tctx,enum FrameType ftype)
969 {
970 int block_size;
971 const ModeTab *mtab = tctx->mtab;
972 int size = tctx->avctx->channels*mtab->fmode[ftype].sub;
973 int16_t *tmp_perm = (int16_t *) tctx->tmp_buf;
974
975 if (ftype == FT_PPC) {
976 size = tctx->avctx->channels;
977 block_size = mtab->ppc_shape_len;
978 } else
979 block_size = mtab->size / mtab->fmode[ftype].sub;
980
981 permutate_in_line(tmp_perm, tctx->n_div[ftype], size,
982 block_size, tctx->length[ftype],
983 tctx->length_change[ftype], ftype);
984
985 transpose_perm(tctx->permut[ftype], tmp_perm, tctx->n_div[ftype],
986 tctx->length[ftype], tctx->length_change[ftype]);
987
988 linear_perm(tctx->permut[ftype], tctx->permut[ftype], size,
989 size*block_size);
990 }
991
992 static av_cold void init_bitstream_params(TwinContext *tctx)
993 {
994 const ModeTab *mtab = tctx->mtab;
995 int n_ch = tctx->avctx->channels;
996 int total_fr_bits = tctx->avctx->bit_rate*mtab->size/
997 tctx->avctx->sample_rate;
998
999 int lsp_bits_per_block = n_ch*(mtab->lsp_bit0 + mtab->lsp_bit1 +
1000 mtab->lsp_split*mtab->lsp_bit2);
1001
1002 int ppc_bits = n_ch*(mtab->pgain_bit + mtab->ppc_shape_bit +
1003 mtab->ppc_period_bit);
1004
1005 int bsize_no_main_cb[3];
1006 int bse_bits[3];
1007 int i;
1008 enum FrameType frametype;
1009
1010 for (i = 0; i < 3; i++)
1011 // +1 for history usage switch
1012 bse_bits[i] = n_ch *
1013 (mtab->fmode[i].bark_n_coef * mtab->fmode[i].bark_n_bit + 1);
1014
1015 bsize_no_main_cb[2] = bse_bits[2] + lsp_bits_per_block + ppc_bits +
1016 WINDOW_TYPE_BITS + n_ch*GAIN_BITS;
1017
1018 for (i = 0; i < 2; i++)
1019 bsize_no_main_cb[i] =
1020 lsp_bits_per_block + n_ch*GAIN_BITS + WINDOW_TYPE_BITS +
1021 mtab->fmode[i].sub*(bse_bits[i] + n_ch*SUB_GAIN_BITS);
1022
1023 // The remaining bits are all used for the main spectrum coefficients
1024 for (i = 0; i < 4; i++) {
1025 int bit_size;
1026 int vect_size;
1027 int rounded_up, rounded_down, num_rounded_down, num_rounded_up;
1028 if (i == 3) {
1029 bit_size = n_ch * mtab->ppc_shape_bit;
1030 vect_size = n_ch * mtab->ppc_shape_len;
1031 } else {
1032 bit_size = total_fr_bits - bsize_no_main_cb[i];
1033 vect_size = n_ch * mtab->size;
1034 }
1035
1036 tctx->n_div[i] = (bit_size + 13) / 14;
1037
1038 rounded_up = (bit_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1039 rounded_down = (bit_size )/tctx->n_div[i];
1040 num_rounded_down = rounded_up * tctx->n_div[i] - bit_size;
1041 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1042 tctx->bits_main_spec[0][i][0] = (rounded_up + 1)/2;
1043 tctx->bits_main_spec[1][i][0] = (rounded_up )/2;
1044 tctx->bits_main_spec[0][i][1] = (rounded_down + 1)/2;
1045 tctx->bits_main_spec[1][i][1] = (rounded_down )/2;
1046 tctx->bits_main_spec_change[i] = num_rounded_up;
1047
1048 rounded_up = (vect_size + tctx->n_div[i] - 1)/tctx->n_div[i];
1049 rounded_down = (vect_size )/tctx->n_div[i];
1050 num_rounded_down = rounded_up * tctx->n_div[i] - vect_size;
1051 num_rounded_up = tctx->n_div[i] - num_rounded_down;
1052 tctx->length[i][0] = rounded_up;
1053 tctx->length[i][1] = rounded_down;
1054 tctx->length_change[i] = num_rounded_up;
1055 }
1056
1057 for (frametype = FT_SHORT; frametype <= FT_PPC; frametype++)
1058 construct_perm_table(tctx, frametype);
1059 }
1060
1061 static av_cold int twin_decode_init(AVCodecContext *avctx)
1062 {
1063 TwinContext *tctx = avctx->priv_data;
1064 int isampf = avctx->sample_rate/1000;
1065 int ibps = avctx->bit_rate/(1000 * avctx->channels);
1066
1067 tctx->avctx = avctx;
1068 avctx->sample_fmt = SAMPLE_FMT_FLT;
1069
1070 if (avctx->channels > 2) {
1071 av_log(avctx, AV_LOG_ERROR, "Unsupported number of channels: %i\n",
1072 avctx->channels);
1073 return -1;
1074 }
1075
1076 switch ((isampf << 8) + ibps) {
1077 case (8 <<8) + 8: tctx->mtab = &mode_08_08; break;
1078 case (11<<8) + 8: tctx->mtab = &mode_11_08; break;
1079 case (11<<8) + 10: tctx->mtab = &mode_11_10; break;
1080 case (16<<8) + 16: tctx->mtab = &mode_16_16; break;
1081 case (22<<8) + 20: tctx->mtab = &mode_22_20; break;
1082 case (22<<8) + 24: tctx->mtab = &mode_22_24; break;
1083 case (22<<8) + 32: tctx->mtab = &mode_22_32; break;
1084 case (44<<8) + 40: tctx->mtab = &mode_44_40; break;
1085 case (44<<8) + 48: tctx->mtab = &mode_44_48; break;
1086 default:
1087 av_log(avctx, AV_LOG_ERROR, "This version does not support %d kHz - %d kbit/s/ch mode.\n", isampf, isampf);
1088 return -1;
1089 }
1090
1091 dsputil_init(&tctx->dsp, avctx);
1092 init_mdct_win(tctx);
1093 init_bitstream_params(tctx);
1094
1095 memset_float(tctx->bark_hist[0][0], 0.1, FF_ARRAY_ELEMS(tctx->bark_hist));
1096
1097 return 0;
1098 }
1099
1100 static av_cold int twin_decode_close(AVCodecContext *avctx)
1101 {
1102 TwinContext *tctx = avctx->priv_data;
1103 int i;
1104
1105 for (i = 0; i < 3; i++) {
1106 ff_mdct_end(&tctx->mdct_ctx[i]);
1107 av_free(tctx->cos_tabs[i]);
1108 }
1109
1110
1111 av_free(tctx->curr_frame);
1112 av_free(tctx->spectrum);
1113 av_free(tctx->prev_frame);
1114 av_free(tctx->tmp_buf);
1115
1116 return 0;
1117 }
1118
1119 AVCodec twinvq_decoder =
1120 {
1121 "twinvq",
1122 AVMEDIA_TYPE_AUDIO,
1123 CODEC_ID_TWINVQ,
1124 sizeof(TwinContext),
1125 twin_decode_init,
1126 NULL,
1127 twin_decode_close,
1128 twin_decode_frame,
1129 .long_name = NULL_IF_CONFIG_SMALL("VQF TwinVQ"),
1130 };